Ask any maintenance manager, distributor, or aftermarket team where their working capital goes, and the answer is often the same: it is sitting on a shelf. Spare parts and slow-moving SKUs are notoriously hard to plan. They sell in ones and twos, with long stretches of nothing in between, and then — without warning — three units go out the door in a single week. This pattern breaks the tools most companies use to manage inventory, and the cost shows up in two places at once: capital frozen in dead stock, and stockouts on the exact part a customer needed today.

The problem spreadsheets can't see

Demand planners describe two properties that make spare parts different from fast-moving goods. The first is intermittence: most time periods have zero demand. The second is lumpiness: when demand does occur, the quantity varies widely. A part might see demand once every six weeks, and when it does, it could be one unit or eight. Statisticians summarize this with two numbers — the average demand interval (ADI), which measures how often demand occurs, and the squared coefficient of variation (CV²) of the non-zero demand sizes, which measures how variable those quantities are. High values on both is the classic "lumpy" quadrant, and it is where conventional forecasting quietly falls apart.

Why min/max rules and moving averages fail

Most ERP systems and spreadsheets forecast with a moving average or simple exponential smoothing, then set a reorder point as average demand over the lead time plus a safety-stock term derived from a normal distribution. Both assumptions are wrong for intermittent parts. A moving average of a series that is zero most of the time produces a small, smooth number that never matches reality: it is always too high right after a sale and too low just before the next one. And the normal distribution — symmetric, centered on the mean — is a poor description of demand that is zero most weeks and occasionally spikes. The real lead-time demand distribution is skewed and discrete, so a safety stock sized from a bell curve systematically misjudges the tail that actually causes stockouts.

The result is the familiar spare-parts paradox: high inventory value and poor service levels at the same time. You are overstocked on the parts that rarely move and understocked on the ones that matter when a line goes down.

What actually works: statistical methods built for intermittence

The foundational technique is Croston's method (1972), which is deceptively simple and still a strong baseline. Instead of smoothing the raw series, it separates demand into two components — the size of demand when it occurs, and the interval between demand occurrences — and applies exponential smoothing to each one independently. Dividing the smoothed size by the smoothed interval gives a demand-rate estimate that does not collapse to noise between sales. Croston's original estimator carries a known positive bias, which the Syntetos-Boylan Approximation (SBA) corrects; for parts whose demand is fading toward obsolescence, the Teunter-Syntetos-Babai (TSB) method updates a demand probability every period and reacts to declining or dying items far better.

Two practices make these methods work in the real world. The first is classification: use the ADI and CV² of each part to route it to the right method, rather than forcing one model on the whole catalogue. Smooth, fast movers can stay on classical exponential smoothing or ARIMA; intermittent and lumpy parts go to Croston-family models. The second is to forecast the full lead-time demand distribution, not just a point estimate. Techniques such as demand bootstrapping resample the historical demand pattern to build an empirical distribution of how much could be needed over a replenishment lead time — exactly the quantity you need to set inventory correctly.

Where machine learning helps — and where it doesn't

Machine learning is not a replacement for statistical methods on spare parts; it is a complement that pays off when there is genuine signal to learn. On a single, isolated, purely random intermittent series, a gradient-boosting model has nothing to grip and a well-tuned Croston/SBA model is hard to beat. The picture changes when demand has drivers you can encode as features: the installed base of the equipment a part serves, scheduled maintenance and warranty cycles, the age of machines in the field, seasonality, promotions, or leading indicators from related parts. Models such as LightGBM can turn those features into calibrated forecasts, and quantile regression lets you predict the specific service-level percentile you care about instead of just an average.

The other place ML earns its keep is scale. When you have thousands of related parts, a single global model trained across the whole catalogue can borrow strength across items — learning shared patterns that a per-part model, starved of data, never could. The right architecture depends on the data, which is why we treat model selection as an empirical question and let a holdout comparison decide, rather than committing to one engine in advance.

From forecast to inventory decisions

A forecast is only useful once it becomes an inventory decision. The bridge is the lead-time demand distribution. Instead of "average demand plus a normal safety factor", the reorder point becomes a quantile of that distribution at your target service level — the on-hand position at which the probability of covering demand until the next delivery meets your fill-rate goal. Safety stock falls out of the same distribution rather than a bell-curve formula, so it is sized for the skewed tail that actually drives stockouts. From there the downstream numbers follow naturally:

  • Reorder points set per part from its own demand distribution and lead time, differentiated by how critical the part is.
  • Safety stock sized to a chosen fill rate, so a critical part gets protection a cheap consumable does not.
  • Projected stockout dates from current on-hand and on-order against the forecast, turning planning from reactive to forward-looking.
  • Freed working capital, by identifying parts that min/max chronically over-orders and flagging obsolete items that should never trigger a reorder at all.

Operationalizing it in production

Knowing the right methods is one thing; running them every day across a live catalogue is another. That is why MLAIA built and operates NextDemand, our spare-parts demand-forecasting product — a platform that classifies every SKU, fits the appropriate statistical or machine-learning model, and turns the resulting distributions into reorder points, safety stock, and projected stockout dates that planners can act on. It is the same approach described here, packaged so a team does not have to assemble it from scratch. If your spare-parts inventory feels simultaneously too expensive and too unreliable, that gap is usually a forecasting problem in disguise — and it is a solvable one.